libvirt/HACKING

-*- buffer-read-only: t -*- vi: set ro:
DO NOT EDIT THIS FILE!  IT IS GENERATED AUTOMATICALLY
from docs/hacking.html.in!



                         Contributor guidelines
                         ======================



General tips for contributing patches
=====================================
(1) Discuss any large changes on the mailing list first. Post patches early and
listen to feedback.

(2) Post patches in unified diff format, with git rename detection enabled. You
need a one-time setup of:

  git config diff.renames true

After that, a command similar to this should work:

  diff -urp libvirt.orig/ libvirt.modified/ > libvirt-myfeature.patch

or:

  git diff > libvirt-myfeature.patch

Also, for code motion patches, you may find that "git diff --patience"
provides an easier-to-read patch. However, the usual workflow of libvirt
developer is:

  git checkout master
  git pull
  git checkout -t origin -b workbranch
  Hack, committing any changes along the way

More hints on compiling can be found here <compiling.html>. When you want to
post your patches:

  git pull --rebase
  (fix any conflicts)
  git send-email --cover-letter --no-chain-reply-to --annotate \
                 --to=libvir-list@redhat.com master

(Note that the "git send-email" subcommand may not be in the main git package
and using it may require installation of a separate package, for example the
"git-email" package in Fedora.) For a single patch you can omit
"--cover-letter", but a series of two or more patches needs a cover letter. If
you get tired of typing "--to=libvir-list@redhat.com" designation you can set
it in git config:

  git config sendemail.to libvir-list@redhat.com

Please follow this as close as you can, especially the rebase and git
send-email part, as it makes life easier for other developers to review your
patch set. One should avoid sending patches as attachments, but rather send
them in email body along with commit message. If a developer is sending
another version of the patch (e.g. to address review comments), he is advised
to note differences to previous versions after the "---" line in the patch so
that it helps reviewers but doesn't become part of git history. Moreover, such
patch needs to be prefixed correctly with "--subject-prefix=PATCHv2" appended
to "git send-email" (substitute "v2" with the correct version if needed
though).



(3) In your commit message, make the summary line reasonably short (60 characters
is typical), followed by a blank line, followed by any longer description of
why your patch makes sense. If the patch fixes a regression, and you know what
commit introduced the problem, mentioning that is useful. If the patch
resolves a bugzilla report, mentioning the URL of the bug number is useful;
but also summarize the issue rather than making all readers follow the link.
You can use 'git shortlog -30' to get an idea of typical summary lines.
Libvirt does not currently attach any meaning to Signed-off-by: lines, so it
is up to you if you want to include or omit them in the commit message.



(4) Split large changes into a series of smaller patches, self-contained if
possible, with an explanation of each patch and an explanation of how the
sequence of patches fits together. Moreover, please keep in mind that it's
required to be able to compile cleanly (*including* "make check" and "make
syntax-check") after each patch. A feature does not have to work until the end
of a series, but intermediate patches must compile and not cause test-suite
failures (this is to preserve the usefulness of "git bisect", among other
things).



(5) Make sure your patches apply against libvirt GIT. Developers only follow GIT
and don't care much about released versions.

(6) Run the automated tests on your code before submitting any changes. In
particular, configure with compile warnings set to -Werror. This is done
automatically for a git checkout; from a tarball, use:

  ./configure --enable-werror

and run the tests:

  make check
  make syntax-check
  make -C tests valgrind

Valgrind <http://valgrind.org/> is a test that checks for memory management
issues, such as leaks or use of uninitialized variables.

Some tests are skipped by default in a development environment, based on the
time they take in comparison to the likelihood that those tests will turn up
problems during incremental builds. These tests default to being run when
building from a tarball or with the configure option --enable-expensive-tests;
you can also force a one-time toggle of these tests by setting
VIR_TEST_EXPENSIVE to 0 or 1 at make time, as in:

  make check VIR_TEST_EXPENSIVE=1

If you encounter any failing tests, the VIR_TEST_DEBUG environment variable
may provide extra information to debug the failures. Larger values of
VIR_TEST_DEBUG may provide larger amounts of information:

  VIR_TEST_DEBUG=1 make check    (or)
  VIR_TEST_DEBUG=2 make check

When debugging failures during development, it is possible to focus in on just
the failing subtests by using TESTS and VIR_TEST_RANGE:

  make check VIR_TEST_DEBUG=1 VIR_TEST_RANGE=3-5 TESTS=qemuxml2argvtest

Also, individual tests can be run from inside the "tests/" directory, like:

  ./qemuxml2xmltest

There is also a "./run" script at the top level, to make it easier to run
programs that have not yet been installed, as well as to wrap invocations of
various tests under gdb or Valgrind.



(7) The Valgrind test should produce similar output to "make check". If the output
has traces within libvirt API's, then investigation is required in order to
determine the cause of the issue. Output such as the following indicates some
sort of leak:

==5414== 4 bytes in 1 blocks are definitely lost in loss record 3 of 89
==5414==    at 0x4A0881C: malloc (vg_replace_malloc.c:270)
==5414==    by 0x34DE0AAB85: xmlStrndup (in /usr/lib64/libxml2.so.2.7.8)
==5414==    by 0x4CC97A6: virDomainVideoDefParseXML (domain_conf.c:7410)
==5414==    by 0x4CD581D: virDomainDefParseXML (domain_conf.c:10188)
==5414==    by 0x4CD8C73: virDomainDefParseNode (domain_conf.c:10640)
==5414==    by 0x4CD8DDB: virDomainDefParse (domain_conf.c:10590)
==5414==    by 0x41CB1D: testCompareXMLToArgvHelper (qemuxml2argvtest.c:100)
==5414==    by 0x41E20F: virtTestRun (testutils.c:161)
==5414==    by 0x41C7CB: mymain (qemuxml2argvtest.c:866)
==5414==    by 0x41E84A: virtTestMain (testutils.c:723)
==5414==    by 0x34D9021734: (below main) (in /usr/lib64/libc-2.15.so)

In this example, the "virDomainDefParseXML()" had an error path where the
"virDomainVideoDefPtr video" pointer was not properly disposed. By simply
adding a "virDomainVideoDefFree(video);" in the error path, the issue was
resolved.

Another common mistake is calling a printing function, such as "VIR_DEBUG()"
without initializing a variable to be printed. The following example involved
a call which could return an error, but not set variables passed by reference
to the call. The solution was to initialize the variables prior to the call.

==4749== Use of uninitialised value of size 8
==4749==    at 0x34D904650B: _itoa_word (in /usr/lib64/libc-2.15.so)
==4749==    by 0x34D9049118: vfprintf (in /usr/lib64/libc-2.15.so)
==4749==    by 0x34D9108F60: __vasprintf_chk (in /usr/lib64/libc-2.15.so)
==4749==    by 0x4CAEEF7: virVasprintf (stdio2.h:199)
==4749==    by 0x4C8A55E: virLogVMessage (virlog.c:814)
==4749==    by 0x4C8AA96: virLogMessage (virlog.c:751)
==4749==    by 0x4DA0056: virNetTLSContextCheckCertKeyUsage (virnettlscontext.c:225)
==4749==    by 0x4DA06DB: virNetTLSContextCheckCert (virnettlscontext.c:439)
==4749==    by 0x4DA1620: virNetTLSContextNew (virnettlscontext.c:562)
==4749==    by 0x4DA26FC: virNetTLSContextNewServer (virnettlscontext.c:927)
==4749==    by 0x409C39: testTLSContextInit (virnettlscontexttest.c:467)
==4749==    by 0x40AB8F: virtTestRun (testutils.c:161)

Valgrind will also find some false positives or code paths which cannot be
resolved by making changes to the libvirt code. For these paths, it is
possible to add a filter to avoid the errors. For example:

==4643== 7 bytes in 1 blocks are possibly lost in loss record 4 of 20
==4643==    at 0x4A0881C: malloc (vg_replace_malloc.c:270)
==4643==    by 0x34D90853F1: strdup (in /usr/lib64/libc-2.15.so)
==4643==    by 0x34EEC2C08A: ??? (in /usr/lib64/libnl.so.1.1)
==4643==    by 0x34EEC15B81: ??? (in /usr/lib64/libnl.so.1.1)
==4643==    by 0x34D8C0EE15: call_init.part.0 (in /usr/lib64/ld-2.15.so)
==4643==    by 0x34D8C0EECF: _dl_init (in /usr/lib64/ld-2.15.so)
==4643==    by 0x34D8C01569: ??? (in /usr/lib64/ld-2.15.so)


In this instance, it is acceptable to modify the "tests/.valgrind.supp" file
in order to add a suppression filter. The filter should be unique enough to
not suppress real leaks, but it should be generic enough to cover multiple
code paths. The format of the entry can be found in the documentation found at
the Valgrind home page <http://valgrind.org/>. The following trace was added
to "tests/.valgrind.supp" in order to suppress the warning:

{
    dlInitMemoryLeak1
    Memcheck:Leak
    fun:?alloc
    ...
    fun:call_init.part.0
    fun:_dl_init
    ...
    obj:*/lib*/ld-2.*so*
}



(8) Update tests and/or documentation, particularly if you are adding a new
feature or changing the output of a program.



There is more on this subject, including lots of links to background reading
on the subject, on Richard Jones' guide to working with open source projects
<http://people.redhat.com/rjones/how-to-supply-code-to-open-source-projects/>.


Code indentation
================
Libvirt's C source code generally adheres to some basic code-formatting
conventions. The existing code base is not totally consistent on this front,
but we do prefer that contributed code be formatted similarly. In short, use
spaces-not-TABs for indentation, use 4 spaces for each indentation level, and
other than that, follow the K&R style.

If you use Emacs, the project includes a file .dir-locals.el that sets up the
preferred indentation. If you use vim, append the following to your ~/.vimrc
file:

  set nocompatible
  filetype on
  set autoindent
  set smartindent
  set cindent
  set tabstop=8
  set shiftwidth=4
  set expandtab
  set cinoptions=(0,:0,l1,t0,L3
  filetype plugin indent on
  au FileType make setlocal noexpandtab
  au BufRead,BufNewFile *.am setlocal noexpandtab
  match ErrorMsg /\s\+$\| \+\ze\t/

Or if you don't want to mess your ~/.vimrc up, you can save the above into a
file called .lvimrc (not .vimrc) located at the root of libvirt source, then
install a vim script from
http://www.vim.org/scripts/script.php?script_id=1408, which will load the
.lvimrc only when you edit libvirt code.


Code formatting (especially for new code)
=========================================
With new code, we can be even more strict. Please apply the following function
(using GNU indent) to any new code. Note that this also gives you an idea of
the type of spacing we prefer around operators and keywords:

  indent-libvirt()
  {
    indent -bad -bap -bbb -bli4 -br -ce -brs -cs -i4 -l75 -lc75 \
      -sbi4 -psl -saf -sai -saw -sbi4 -ss -sc -cdw -cli4 -npcs -nbc \
      --no-tabs "$@"
  }

Note that sometimes you'll have to post-process that output further, by piping
it through "expand -i", since some leading TABs can get through. Usually
they're in macro definitions or strings, and should be converted anyhow.

Libvirt requires a C99 compiler for various reasons. However, most of the code
base prefers to stick to C89 syntax unless there is a compelling reason
otherwise. For example, it is preferable to use "/* */" comments rather than
"//". Also, when declaring local variables, the prevailing style has been to
declare them at the beginning of a scope, rather than immediately before use.


Bracket spacing
===============
The keywords "if", "for", "while", and "switch" must have a single space
following them before the opening bracket. E.g.

      if(foo)   // Bad
      if (foo)  // Good

Function implementations mustnothave any whitespace between the function name and the opening bracket. E.g.

      int foo (int wizz)  // Bad
      int foo(int wizz)   // Good

Function calls mustnothave any whitespace between the function name and the opening bracket. E.g.

      bar = foo (wizz);  // Bad
      bar = foo(wizz);   // Good

Function typedefs mustnothave any whitespace between the closing bracket of the function name and
opening bracket of the arg list. E.g.

      typedef int (*foo) (int wizz);  // Bad
      typedef int (*foo)(int wizz);   // Good

There must not be any whitespace immediately following any opening bracket, or
immediately prior to any closing bracket. E.g.

      int foo( int wizz );  // Bad
      int foo(int wizz);    // Good


Commas
======
Commas should always be followed by a space or end of line, and never have
leading space; this is enforced during 'make syntax-check'.

      call(a,b ,c);// Bad
      call(a, b, c); // Good

When declaring an enum or using a struct initializer that occupies more than
one line, use a trailing comma. That way, future edits to extend the list only
have to add a line, rather than modify an existing line to add the
intermediate comma. Any sentinel enumerator value with a name ending in _LAST
is exempt, since you would extend such an enum before the _LAST element.
Another reason to favor trailing commas is that it requires less effort to
produce via code generators. Note that the syntax checker is unable to enforce
a style of trailing commas, so there are counterexamples in existing code
which do not use it; also, while C99 allows trailing commas, remember that
JSON and XDR do not.

      enum {
          VALUE_ONE,
          VALUE_TWO // Bad
      };
      enum {
          VALUE_THREE,
          VALUE_FOUR, // Good
      };


Semicolons
==========
Semicolons should never have a space beforehand. Inside the condition of a
"for" loop, there should always be a space or line break after each semicolon,
except for the special case of an infinite loop (although more infinite loops
use "while"). While not enforced, loop counters generally use post-increment.

      for (i = 0 ;i < limit ; ++i) { // Bad
      for (i = 0; i < limit; i++) { // Good
      for (;;) { // ok
      while (1) { // Better

Empty loop bodies are better represented with curly braces and a comment,
although use of a semicolon is not currently rejected.

      while ((rc = waitpid(pid, &st, 0) == -1) &&
             errno == EINTR); // ok
      while ((rc = waitpid(pid, &st, 0) == -1) &&
             errno == EINTR) { // Better
          /* nothing */
      }


Curly braces
============
Omit the curly braces around an "if", "while", "for" etc. body only when that
body occupies a single line. In every other case we require the braces. This
ensures that it is trivially easy to identify a single-'statement' loop: each
has only one 'line' in its body.

Omitting braces with a single-line body is fine:

  while (expr) // one-line body -> omitting curly braces is ok
      single_line_stmt();

However, the moment your loop/if/else body extends on to a second line, for
whatever reason (even if it's just an added comment), then you should add
braces. Otherwise, it would be too easy to insert a statement just before that
comment (without adding braces), thinking it is already a multi-statement loop:

  while (true) // BAD! multi-line body with no braces
      /* comment... */
      single_line_stmt();

Do this instead:

  while (true) { // Always put braces around a multi-line body.
      /* comment... */
      single_line_stmt();
  }

There is one exception: when the second body line is not at the same
indentation level as the first body line:

  if (expr)
      die("a diagnostic that would make this line"
          " extend past the 80-column limit"));

It is safe to omit the braces in the code above, since the further-indented
second body line makes it obvious that this is still a single-statement body.

To reiterate, don't do this:

  if (expr)            // BAD: no braces around...
      while (expr_2) { // ... a multi-line body
          ...
      }

Do this, instead:

  if (expr) {
      while (expr_2) {
          ...
      }
  }

However, there is one exception in the other direction, when even a one-line
block should have braces. That occurs when that one-line, brace-less block is
an "if" or "else" block, and the counterpart block *does* use braces. In that
case, put braces around both blocks. Also, if the "else" block is much shorter
than the "if" block, consider negating the "if"-condition and swapping the
bodies, putting the short block first and making the longer, multi-line block
be the "else" block.

  if (expr) {
      ...
      ...
  }
  else
      x = y;    // BAD: braceless "else" with braced "then",
                // and short block last

  if (expr)
      x = y;    // BAD: braceless "if" with braced "else"
  else {
      ...
      ...
  }

Keeping braces consistent and putting the short block first is preferred,
especially when the multi-line body is more than a few lines long, because it
is easier to read and grasp the semantics of an if-then-else block when the
simpler block occurs first, rather than after the more involved block:

  if (!expr) {
    x = y; // putting the smaller block first is more readable
  } else {
      ...
      ...
  }

But if negating a complex condition is too ugly, then at least add braces:

  if (complex expr not worth negating) {
      ...
      ...
  } else {
      x = y;
  }


Preprocessor
============
Macros defined with an ALL_CAPS name should generally be assumed to be unsafe
with regards to arguments with side-effects (that is, MAX(a++, b--) might
increment a or decrement b too many or too few times). Exceptions to this rule
are explicitly documented for macros in viralloc.h and virstring.h.

For variadic macros, stick with C99 syntax:

  #define vshPrint(_ctl, ...) fprintf(stdout, __VA_ARGS__)

Use parenthesis when checking if a macro is defined, and use indentation to
track nesting:

  #if defined(HAVE_POSIX_FALLOCATE) && !defined(HAVE_FALLOCATE)
  # define fallocate(a,ignored,b,c) posix_fallocate(a,b,c)
  #endif


C types
=======
Use the right type.

Scalars
-------
- If you're using "int" or "long", odds are good that there's a better type.

- If a variable is counting something, be sure to declare it with an unsigned
type.

- If it's memory-size-related, use "size_t" (use "ssize_t" only if required).

- If it's file-size related, use uintmax_t, or maybe "off_t".

- If it's file-offset related (i.e., signed), use "off_t".

- If it's just counting small numbers use "unsigned int"; (on all but oddball
embedded systems, you can assume that that type is at least four bytes wide).

- If a variable has boolean semantics, give it the "bool" type and use the
corresponding "true" and "false" macros. It's ok to include <stdbool.h>, since
libvirt's use of gnulib ensures that it exists and is usable.

- In the unusual event that you require a specific width, use a standard type
like "int32_t", "uint32_t", "uint64_t", etc.

- While using "bool" is good for readability, it comes with minor caveats:

-- Don't use "bool" in places where the type size must be constant across all
systems, like public interfaces and on-the-wire protocols. Note that it would
be possible (albeit wasteful) to use "bool" in libvirt's logical wire
protocol, since XDR maps that to its lower-level "bool_t" type, which *is*
fixed-size.

-- Don't compare a bool variable against the literal, "true", since a value with
a logical non-false value need not be "1". I.e., don't write "if (seen ==
true) ...". Rather, write "if (seen)...".





Of course, take all of the above with a grain of salt. If you're about to use
some system interface that requires a type like "size_t", "pid_t" or "off_t",
use matching types for any corresponding variables.

Also, if you try to use e.g., "unsigned int" as a type, and that conflicts
with the signedness of a related variable, sometimes it's best just to use the
*wrong* type, if 'pulling the thread' and fixing all related variables would
be too invasive.

Finally, while using descriptive types is important, be careful not to go
overboard. If whatever you're doing causes warnings, or requires casts, then
reconsider or ask for help.

Pointers
--------
Ensure that all of your pointers are 'const-correct'. Unless a pointer is used
to modify the pointed-to storage, give it the "const" attribute. That way, the
reader knows up-front that this is a read-only pointer. Perhaps more
importantly, if we're diligent about this, when you see a non-const pointer,
you're guaranteed that it is used to modify the storage it points to, or it is
aliased to another pointer that is.


Low level memory management
===========================
Use of the malloc/free/realloc/calloc APIs is deprecated in the libvirt
codebase, because they encourage a number of serious coding bugs and do not
enable compile time verification of checks for NULL. Instead of these
routines, use the macros from viralloc.h.

- To allocate a single object:

  virDomainPtr domain;

  if (VIR_ALLOC(domain) < 0)
      return NULL;



- To allocate an array of objects:

  virDomainPtr domains;
  size_t ndomains = 10;

  if (VIR_ALLOC_N(domains, ndomains) < 0)
      return NULL;



- To allocate an array of object pointers:

  virDomainPtr *domains;
  size_t ndomains = 10;

  if (VIR_ALLOC_N(domains, ndomains) < 0)
      return NULL;



- To re-allocate the array of domains to be 1 element longer (however, note that
repeatedly expanding an array by 1 scales quadratically, so this is
recommended only for smaller arrays):

  virDomainPtr domains;
  size_t ndomains = 0;

  if (VIR_EXPAND_N(domains, ndomains, 1) < 0)
      return NULL;
  domains[ndomains - 1] = domain;



- To ensure an array has room to hold at least one more element (this approach
scales better, but requires tracking allocation separately from usage)

  virDomainPtr domains;
  size_t ndomains = 0;
  size_t ndomains_max = 0;

  if (VIR_RESIZE_N(domains, ndomains_max, ndomains, 1) < 0)
      return NULL;
  domains[ndomains++] = domain;



- To trim an array of domains from its allocated size down to the actual used
size:

  virDomainPtr domains;
  size_t ndomains = x;
  size_t ndomains_max = y;

  VIR_SHRINK_N(domains, ndomains_max, ndomains_max - ndomains);



- To free an array of domains:

  virDomainPtr domains;
  size_t ndomains = x;
  size_t ndomains_max = y;
  size_t i;

  for (i = 0; i < ndomains; i++)
      VIR_FREE(domains[i]);
  VIR_FREE(domains);
  ndomains_max = ndomains = 0;






File handling
=============
Usage of the "fdopen()", "close()", "fclose()" APIs is deprecated in libvirt
code base to help avoiding double-closing of files or file descriptors, which
is particularly dangerous in a multi-threaded application. Instead of these
APIs, use the macros from virfile.h

- Open a file from a file descriptor:

  if ((file = VIR_FDOPEN(fd, "r")) == NULL) {
      virReportSystemError(errno, "%s",
                           _("failed to open file from file descriptor"));
      return -1;
  }
  /* fd is now invalid; only access the file using file variable */



- Close a file descriptor:

  if (VIR_CLOSE(fd) < 0) {
      virReportSystemError(errno, "%s", _("failed to close file"));
  }



- Close a file:

  if (VIR_FCLOSE(file) < 0) {
      virReportSystemError(errno, "%s", _("failed to close file"));
  }



- Close a file or file descriptor in an error path, without losing the previous
"errno" value:

  VIR_FORCE_CLOSE(fd);
  VIR_FORCE_FCLOSE(file);






String comparisons
==================
Do not use the strcmp, strncmp, etc functions directly. Instead use one of the
following semantically named macros

- For strict equality:

  STREQ(a,b)
  STRNEQ(a,b)



- For case insensitive equality:

  STRCASEEQ(a,b)
  STRCASENEQ(a,b)



- For strict equality of a substring:

  STREQLEN(a,b,n)
  STRNEQLEN(a,b,n)



- For case insensitive equality of a substring:

  STRCASEEQLEN(a,b,n)
  STRCASENEQLEN(a,b,n)



- For strict equality of a prefix:

  STRPREFIX(a,b)



- To avoid having to check if a or b are NULL:

  STREQ_NULLABLE(a, b)
  STRNEQ_NULLABLE(a, b)






String copying
==============
Do not use the strncpy function. According to the man page, it does *not*
guarantee a NULL-terminated buffer, which makes it extremely dangerous to use.
Instead, use one of the functionally equivalent functions:

  virStrncpy(char *dest, const char *src, size_t n, size_t destbytes)

The first three arguments have the same meaning as for strncpy; namely the
destination, source, and number of bytes to copy, respectively. The last
argument is the number of bytes available in the destination string; if a copy
of the source string (including a \0) will not fit into the destination, no
bytes are copied and the routine returns NULL. Otherwise, n bytes from the
source are copied into the destination and a trailing \0 is appended.

  virStrcpy(char *dest, const char *src, size_t destbytes)

Use this variant if you know you want to copy the entire src string into dest.
Note that this is a macro, so arguments could be evaluated more than once.
This is equivalent to virStrncpy(dest, src, strlen(src), destbytes)

  virStrcpyStatic(char *dest, const char *src)

Use this variant if you know you want to copy the entire src string into dest
*and* you know that your destination string is a static string (i.e. that
sizeof(dest) returns something meaningful). Note that this is a macro, so
arguments could be evaluated more than once. This is equivalent to
virStrncpy(dest, src, strlen(src), sizeof(dest)).

  VIR_STRDUP(char *dst, const char *src);
  VIR_STRNDUP(char *dst, const char *src, size_t n);

You should avoid using strdup or strndup directly as they do not report
out-of-memory error, and do not allow a NULL source. Use VIR_STRDUP or
VIR_STRNDUP macros instead, which return 0 for NULL source, 1 for successful
copy, and -1 for allocation failure with the error already reported. In very
specific cases, when you don't want to report the out-of-memory error, you can
use VIR_STRDUP_QUIET or VIR_STRNDUP_QUIET, but such usage is very rare and
usually considered a flaw.


Variable length string buffer
=============================
If there is a need for complex string concatenations, avoid using the usual
sequence of malloc/strcpy/strcat/snprintf functions and make use of the
virBuffer API described in virbuffer.h

Typical usage is as follows:

  char *
  somefunction(...)
  {
     virBuffer buf = VIR_BUFFER_INITIALIZER;

     ...

     virBufferAddLit(&buf, "<domain>\n");
     virBufferAsprintf(&buf, "  <memory>%d</memory>\n", memory);
     ...
     virBufferAddLit(&buf, "</domain>\n");

     ...

     if (virBufferCheckError(&buf) < 0)
         return NULL;

     return virBufferContentAndReset(&buf);
  }


Include files
=============
There are now quite a large number of include files, both libvirt internal and
external, and system includes. To manage all this complexity it's best to
stick to the following general plan for all *.c source files:

  /*
   * Copyright notice
   * ....
   * ....
   * ....
   *
   */

  #include <config.h>             Must come first in every file.

  #include <stdio.h>              Any system includes you need.
  #include <string.h>
  #include <limits.h>

  #if WITH_NUMACTL                Some system includes aren't supported
  # include <numa.h>              everywhere so need these #if guards.
  #endif

  #include "internal.h"           Include this first, after system includes.

  #include "util.h"               Any libvirt internal header files.
  #include "buf.h"

  static int
  myInternalFunc()                The actual code.
  {
      ...

Of particular note: *Do not* include libvirt/libvirt.h, libvirt/virterror.h,
libvirt/libvirt-qemu.h, or libvirt/libvirt-lxc.h. They are included by
"internal.h" already and there are some special reasons why you cannot include
these files explicitly. One of the special cases, "libvirt/libvirt.h" is
included prior to "internal.h" in "remote_protocol.x", to avoid exposing
*_LAST enum elements.


Printf-style functions
======================
Whenever you add a new printf-style function, i.e., one with a format string
argument and following "..." in its prototype, be sure to use gcc's printf
attribute directive in the prototype. For example, here's the one for
virAsprintf, in util.h:

  int virAsprintf(char **strp, const char *fmt, ...)
      ATTRIBUTE_FORMAT(printf, 2, 3);

This makes it so gcc's -Wformat and -Wformat-security options can do their
jobs and cross-check format strings with the number and types of arguments.

When printing to a string, consider using virBuffer for incremental
allocations, virAsprintf for a one-shot allocation, and snprintf for
fixed-width buffers. Do not use sprintf, even if you can prove the buffer
won't overflow, since gnulib does not provide the same portability guarantees
for sprintf as it does for snprintf.


Use of goto
===========
The use of goto is not forbidden, and goto is widely used throughout libvirt.
While the uncontrolled use of goto will quickly lead to unmaintainable code,
there is a place for it in well structured code where its use increases
readability and maintainability. In general, if goto is used for error
recovery, it's likely to be ok, otherwise, be cautious or avoid it all
together.

The typical use of goto is to jump to cleanup code in the case of a long list
of actions, any of which may fail and cause the entire operation to fail. In
this case, a function will have a single label at the end of the function.
It's almost always ok to use this style. In particular, if the cleanup code
only involves free'ing memory, then having multiple labels is overkill.
VIR_FREE() and every function named XXXFree() in libvirt is required to handle
NULL as its arg. Thus you can safely call free on all the variables even if
they were not yet allocated (yes they have to have been initialized to NULL).
This is much simpler and clearer than having multiple labels.

There are a couple of signs that a particular use of goto is not ok:

- You're using multiple labels. If you find yourself using multiple labels,
you're strongly encouraged to rework your code to eliminate all but one of
them.

- The goto jumps back up to a point above the current line of code being
executed. Please use some combination of looping constructs to re-execute code
instead; it's almost certainly going to be more understandable by others. One
well-known exception to this rule is restarting an i/o operation following
EINTR.

- The goto jumps down to an arbitrary place in the middle of a function followed
by further potentially failing calls. You should almost certainly be using a
conditional and a block instead of a goto. Perhaps some of your function's
logic would be better pulled out into a helper function.



Although libvirt does not encourage the Linux kernel wind/unwind style of
multiple labels, there's a good general discussion of the issue archived at
KernelTrap <http://kerneltrap.org/node/553/2131>

When using goto, please use one of these standard labels if it makes sense:

      error: A path only taken upon return with an error code
    cleanup: A path taken upon return with success code + optional error
  no_memory: A path only taken upon return with an OOM error code
      retry: If needing to jump upwards (e.g., retry on EINTR)

Top-level labels should be indented by one space (putting them on the
beginning of the line confuses function context detection in git):

int foo()
{
    /* ... do stuff ... */
 cleanup:
    /* ... do other stuff ... */
}


Libvirt committer guidelines
============================
The AUTHORS files indicates the list of people with commit access right who
can actually merge the patches.

The general rule for committing a patch is to make sure it has been reviewed
properly in the mailing-list first, usually if a couple of people gave an ACK
or +1 to a patch and nobody raised an objection on the list it should be good
to go. If the patch touches a part of the code where you're not the main
maintainer, or where you do not have a very clear idea of how things work,
it's better to wait for a more authoritative feedback though. Before
committing, please also rebuild locally, run 'make check syntax-check', and
make sure you don't raise errors. Try to look for warnings too; for example,
configure with

  --enable-compile-warnings=error

which adds -Werror to compile flags, so no warnings get missed

An exception to 'review and approval on the list first' is fixing failures to
build:

- if a recently committed patch breaks compilation on a platform or for a given
driver, then it's fine to commit a minimal fix directly without getting the
review feedback first

- if make check or make syntax-check breaks, if there is an obvious fix, it's
fine to commit immediately. The patch should still be sent to the list (or
tell what the fix was if trivial), and 'make check syntax-check' should pass
too, before committing anything

- fixes for documentation and code comments can be managed in the same way, but
still make sure they get reviewed if non-trivial.